When researchers at New Mexico’s Los Alamos National Laboratory, one of the country’s most important national security labs, were looking to study how to forecast dangerous infectious diseases like dengue, they knew they had a few tools at their disposal.
There was more slow-moving data, like that from the census or examinations of the spread of climate change. And there was the ability to look at social media signals, such as Google Health Trend searches.
The idea, recalls Sara Del Valle, a leader of LANL’s epidemiological forecasting team, is that people often go online to check symptoms they may be feeling before they visit a doctor. “Because dengue is one of the common diseases in Brazil,” Del Valle says, “we could see a lot of interest in dengue on Google, [people searching] about 25 different terms, like ‘mosquito,’ ‘dengue,’ the names of mosquitoes,” and so on.
But relying on these signals isn’t enough to come up with meaningful forecasting models, the kinds that can allow public health officials to issue warnings or urge people in affected areas to take precautions. There’s not enough data and it’s not necessarily accurate enough, LANL officials felt.
For LANL, forecasting infectious diseases like influenza, HIV, dengue, and others is a matter of national security. “They could be used for bio-terrorist purposes,” says Del Valle, “and if enough people get sick, there won’t be enough people for critical positions [in places like] telecom companies, or labs.” As well, says Amanda Ziemann, a remote sensing scientist in LANL’s Space Data Science and Systems group, the spread of infectious diseases around the world can lead to regional political instability.
The need for better forecasting systems is why the researchers realized that they needed to add more data to the mix, and “that’s when we took a detour and started to explore including remote sensing,” says Ziemann.
The best remote sensing technology, they felt, was satellite imagery. And the best partner to work with to provide access to such imagery, was Santa Fe, New Mexico-based predictive intelligence company Descartes Labs, itself a three-year-old spinoff of LANL. Founded to analyze public domain satellite imagery like that available from NASA, the European Space Agency, and others, Descartes got started producing lucrative models of corn crop yields, and soon had built a roster of corporate and government clients, as well as a cloud-based platform that its clients could use to process and analyze data for their own purposes.
As Descartes has grown–it’s now up to 70 employees–it has taken on more of what it calls “impact science.” CEO Mark Johnson explains that its systems could be used for things like predicting future food shortages that might lead to political unrest and “try to send in humanitarian aid before you have to send in military aid five years later.”
Descartes is also currently working on localized experiments–one such project is to build systems that monitor real-time infrared satellite imagery of the forests in New Mexico for outbreaks of fire, and proactively call in first responders. Such a system has the potential to get firefighters to a blaze well before any other method, says Descartes applied scientist lead Caitlin Kontgis.
A proxy measure
For LANL folks like Del Valle and Ziemann, satellites are ideal for determining, at scale and in close-to-real-time, things like vegetation health, climate-based changes in temperature and precipitation, and urbanization. Those are all keys for helping to spot potential hot zones for mosquito-borne diseases like dengue. And so, working with Descartes, LANL set out to see if it could come up with reliable forecasts for dengue in Brazil.
Of course, the more precipitation there is, the more standing water, and thus more mosquitoes. And if you know there are mosquitoes, you can predict that dengue may soon follow the two-week lifespan of the many newly-hatched mosquitoes that proliferate in areas of standing water. But because it’s not easy to see such water in satellite imagery.
“I learned that standing water is the biggest indicator” for mosquitos, and thus dengue, Ziemann says. “If you want to measure standing water in Brazil, it’s hard, but you have a proxy measure–healthy vegetation” that’s much easier to see in satellite imagery.
In an ideal world, epidemiologists would measure mosquito density on the ground, because if you know where the little flying parasites are, especially in relation to people, you could build disease forecasting models, explains Geoffrey Fairchild, who works in information systems and modeling at LANL. The problem, Fairchild adds, is that it’s generally not possible to set mosquito traps at scale, so such a method is useful only for small case studies, particularly in a huge country like Brazil, the world’s fifth-largest by land mass, which is also ground zero for other mosquito-borne diseases like zika and chikungunya. Yet Brazil is also fertile ground for studying such maladies because of its demographics. “It’s a nice test bed,” Fairchild says, “a whole gamut of really highly-developed regions to not developed at all.”
The knowledge of Brazil’s demographics and limitations is what led the researchers to explore ways to quantify secondary measures like instances of healthy vegetation, and thus standing water–in other words, by using Descartes’s systems.
NASA’s ‘Prescient move’
One of Descartes’s foundations was that NASA had been consistently taking and saving satellite imagery of the entire planet–albeit on film, because there wasn’t enough computer storage–since 1972.
That prescient move by NASA, coupled with the fact that dengue has been around in Brazil for quite awhile, makes that country ideal for this kind of research. Other diseases are too new to allow for the kind of historical exploration that could lead the researchers to link what they can see in past satellite imagery to known instances of outbreak. “Dengue seemed to be perfect,” Del Valle says.
The LANL team was working with seven years of historical data–clinical, surveillance, climate, social media, and demographic–things such as the link between poverty and the lack of access to clean water or proper sewage. It put all that data into mathematical models that were then able to express which data sources contributed most to a forecast, and how much they were leading or lagging indicators. Del Valle says that the existence of healthy vegetation provided as much as five weeks of advance warning of a dengue outbreak. By comparison, Google’s now defunct public health trends system showed a two-week lag between increases in searches and dengue outbreaks.
Descartes gave the team the ability to correlate that seven years of data with corresponding historical satellite imagery collected from NASA, the European Space Agency, and the U.S. Geological Survey. LANL was able to collect more than two-dozen terabytes of data from the various kinds of satellite systems used by each agency, and then do complex computing analysis on Descartes’s cloud-based platform.
Further, LANL was able to specify the areas it wanted to study, and Descartes could provide the imagery solely from those locations.
Experimental, for now
Ultimately, LANL doesn’t imagine it’s anywhere close to coming up with a global disease forecasting system, but the Brazil project will ideally lead to increased funding that it can use to analyze additional regions of the world, and to forecast different diseases across the globe.
It also hopes to build an app-based platform that everyone from public health officials to decision makers to the general public can use to see its disease forecasts. Del Valle knows that such a system could have a materially negative impact on the data, because if people think there’s going to be an outbreak, they’ll change their behavior–starting to use more mosquito repellant, for example, or use condoms in areas where diseases are spread through sexual activity. But of course, the ultimate goal is to keep people healthy, not collect data.
When one hears the word “forecast,” it’s tempting to think of the weather, and the major advances in that area over the years. But as the LANL team explained in a recent paper, disease forecasting is far messier than predicting the weather, so bringing satellite-based data into the equation has been a major boon. “Weather is purely judged on physics,” Fairchild says. “If you knew the complete state of the world, you could predict the weather [perfectly]. But the human component [of disease] makes things so much more challenging than weather. It’s not physics. It’s much harder.”